From capybaraa
Scan the whole repository against capybaraa's rules: lean, ask, optimal, terse, clean, sync. A ranked, codebase-wide report of over-engineering, dead code, bad complexity, filler comments, missing validation, and docs/tests out of sync with the code. One line per finding, biggest impact first, lists only, does not apply fixes. Use when the user says "audit this repo against the rules", "capybaraa audit", "/capybaraa-audit", "what can I delete from this repo", or "find the bloat".
How this skill is triggered — by the user, by Claude, or both
Slash command
/capybaraa:capybaraa-auditThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Scan the entire repository against capybaraa's rules and report what a calm senior dev
Scan the entire repository against capybaraa's rules and report what a calm senior dev
would cut or fix. This is the whole-repo counterpart to /capybaraa-review (which only
looks at the current diff). It is a quality pass, not a correctness audit, and not a
security audit: for runtime bugs use /code-review, for exploits use /security-review.
Detailed guidance: references/principles.md.
Map the repo first: source layout, entry points, the dependency manifest. Skip
vendored code, lockfiles, build output, and .git.
Read the real source. For every problem, emit ONE line:
path:line: <tag> <what's wrong>. <the fix>.
Rank the findings by impact, biggest first: the abstraction nobody uses, the dependency a few lines would replace, the dead module, before the stray comment.
Group nothing, pad nothing. End with a one-line verdict naming the top one to three things to delete or fix.
ask: a spec was guessed where nothing said it: a flag, a branch, a config key
built on an assumption. Name the assumption.lean: something that does not need to exist: an abstraction with one caller, a
config for a constant, a dependency for what stdlib or a few lines already do, dead
flexibility kept "for later". Name what to delete or inline.optimal: wrong data structure or needless O(n^2) on a hot path. Name the better one.terse: filler comments that restate the code, dead prose, docs that lie about the
code. Name the lines to cut.clean: dead code, an orphaned old version left beside its replacement, a stale
comment, or an unsanitized input at a trust boundary. Name what to remove or guard.sync: a doc, README, comment, test, or version string that describes a shape the code
no longer has: a renamed symbol, a removed flag, a lagging version. Name what to update
or delete so the repo stops lying about itself.src/cache/CacheManager.js:1: lean: 180-line cache class wraps one Map with no eviction. Replace with a Map, add eviction when it's actually needed.
src/util/dates.js:1: lean: moment imported for one format() call. Use Intl.DateTimeFormat, drop the dependency.
src/legacy/parseV1.js:1: clean: whole module dead since v2 parser landed, no callers. Delete it.
src/api/handler.js:55: clean: req.body.id passed straight into the query, no validation. Guard it at the boundary.
src/report/build.js:30: optimal: findUser called inside the row loop, O(n^2). Build a Map of users once.
README.md:12: sync: documents a --verbose flag removed in v3, no longer parsed. Delete the line.
verdict: delete CacheManager and parseV1, drop moment.
Lists findings only, does not edit files. Over-engineering, mess, dead code, and docs out
of sync are in scope. Correctness bugs, security exploits, and performance profiling are
not, that is what /code-review and /security-review are for. If you spot a real
security hole while reading, flag it in one line and tell the user to run
/security-review. Do not fix it silently.
npx claudepluginhub katipally/capybaraa --plugin capybaraaProvides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.